Three cards from three operators. One finding about AI tool failure modes. Two findings about how AI-era changes are misread by practitioners working from pre-AI assumptions.
Extended reasoning failure, max thinking burns the token budget silently
Simon Willison published a hands-on report on September 22, 2026, testing the newly released Claude Opus 5.5 alongside GPT-6 Sol and GPT-6 Luna. The pricing findings were significant. The most operationally urgent finding was a failure mode most developers haven't hit yet.
Max extended thinking exhausts the full 128K output token budget on routine prompts and returns nothing documents it precisely. Willison ran Opus 5.5 at "max" extended thinking with a prompt to draw an SVG pelican. The model consumed all 128,000 output tokens in pure reasoning and returned nothing. Both test runs failed the same way. Total cost: $2.56 per attempt. Duration: nearly 20 minutes each. The mechanism is structural. Extended thinking allocates thinking tokens from the same pool as output tokens. With "max" enabled and no explicit budget cap, the model can exhaust the entire 128K limit in reasoning before generating a single character of output. The fix is setting a concrete thinking_budget in the API request. The "max" setting, Willison concluded, is effectively unusable until explicit caps are built in.
"Opus 5.5 has a 128,000 maximum output token limit (as do the other Claude models), and it hit that while it was still reasoning about the SVG!"
"This makes me suspect that 'max' is effectively useless — if it over-thinks to breaking point on a stupid SVG prompt I don't trust it not to do the same for more interesting work."
- Simon Willison, Max extended thinking exhausts the full 128K output token budget on routine prompts and returns nothing. Max extended thinking can exhaust the full output token budget on open-ended prompts and return nothing, at $2.56 per failed attempt.
Search quality and cost, the floor that production economics used to provide
Two independent operators published on the same underlying problem: the implicit quality controls that existed before LLMs lowered the cost of content production to near-zero.
Lily Ray posted on September 16, 2026, framing the issue as an economic progression. Programmatic SEO lowered the per-page production cost. LLMs removed it entirely. The economic friction of content production was the last quality gate in search; LLMs removed it captures her core observation. Production cost acted as a quality floor not because publishers chose quality but because mass-producing low-quality content required capital, staff, and sustained operational effort. That friction kept the index from spam saturation. The friction is now gone. Google's algorithmic spam detection is the only remaining gate, without the production-cost barrier that used to constrain supply.
"The marginal cost of publishing is one of the only things standing between a healthy index and total chaos (spammy search results)"
Aleyda Solis published on September 15, 2026, addressing a related misdiagnosis afflicting ecommerce teams. Ecommerce YOY organic traffic drops despite stable ranking positions diagnose SERP layout displacement, not ranking failure identifies the pattern: sites see non-branded organic traffic down 30-50% year-over-year while their rank positions look stable or improved. The instinct is to chase higher rankings. The actual cause is that Google's SERP layouts changed substantially between September 2025 and September 2026. AI Overviews, AI Mode panels, and Shopping packs now occupy the above-the-fold real estate that organic results previously held. The ranking position is the same; the visible screen space assigned to it shrank. A higher rank on a query where AI Mode now answers intent inline does not recover traffic. Rethinking which queries to optimize for, given the new layout, does.
"Are you an ecommerce seeing important YOY drops in organic search traffic from your top performing non-branded queries/topics even if you haven't actually decreased in real positions?"
These two cards connect. Ray explains why the index is flooding with content. Solis explains why operators are mis-measuring its effect on their own traffic and applying the wrong fix.
- Lily Ray, The economic friction of content production was the last quality gate in search; LLMs removed it. LLMs removed the production-cost barrier that previously kept content spam constrained; algorithmic detection is now the only quality gate.
- Aleyda Solis, Ecommerce YOY organic traffic drops despite stable ranking positions diagnose SERP layout displacement, not ranking failure. Ecommerce YOY organic traffic drops despite stable rank positions diagnose SERP layout displacement, not a ranking problem.